MIT used a small-data AI loop to make mRNA vaccines survive room temperature for a year

mRNA vaccines carried by lipid nanoparticles must be stored at -20 to -80°C, which keeps them out of regions without cold-chain infrastructure. At MIT's Koch Institute, Ana Jaklenec and Robert Langer's group had already built polymer-stabilized LNPs that tolerate heat, but those changed the FDA-approved formulations used by Moderna and Pfizer. Trying to stabilize the approved formulations by hand, they screened excipients — sugars, salts, polymers — for months. "Nothing would get us to 100 percent stability," postdoc Khanh Tran said.

The fix was a machine-learning algorithm developed with MIT CSAIL's Mina Konaković Luković, built to predict from very small datasets. The team measured how well each of nearly 50 FDA-approved excipients protected mRNA inside LNPs, chose the five most promising, and let the algorithm predict stabilizing ratios. They tested two formulations at a time in cells and fed results back into the model; after a handful of iterations — a few weeks, not months — one formulation stood out.

Vacuum-dried Covid-19 mRNA vaccines in the new formulation were stored at 37°C for two months or at room temperature for a year, and mice vaccinated afterward mounted immune responses equivalent to a Moderna-style LNP. The same five excipients in a different ratio stabilized a Pfizer-style LNP, and the formulation also worked in solid microneedle patches, which produced immune responses similar to injections. The work appeared in Nature Biotechnology and was funded in part by the Gates Foundation.

Run an active-learning loop over 50 already-FDA-approved excipients: predict ratios, test two at a time in cells, feed results back — converged in weeks.

The cold chain is the bottleneck: -20 to -80°C storage keeps RNA vaccines away from regions without freezers.

Restricting the search to FDA-approved excipients leaves a regulatory path instead of requiring new chemistry.

The algorithm is built for small datasets, matching a lab that cannot run thousands of experiments.

The two-at-a-time feedback loop cut the search from months of manual screening to a few weeks.

When brute-force screening stalls, don't invent new ingredients — search smarter over already-approved ones. Small-data optimization turns months of dead ends into a handful of targeted experiments.

As of the September 2026 announcement the results were preclinical: mouse immune responses and microneedle-patch delivery, not human trials or regulatory review. The researchers say each heat-resistant formulation could carry any mRNA payload, broadening use to therapeutics, controlled-release particles and microneedle patches that require solid-state or heat-stable formulations.